Spatial Autocorrelation

There are good reasons to check for spatial dependencies in our data. Spatial autocorre-lation induced by strategic interaction could be responsible for possible (dis-)similarities between grants awarded to neighboring counties. One reason for suspecting such effects is yardstick competition. We would then expect spending between neighbors to be positively correlated. As yardstick competition essentially requires two governments that strategi-cally interact, this may not be all that relevant in our setting, where the governor decides

68CA stations are low power TV stations which are given protected status by the FCC because they convey local information. We exclude stations above 10kw, because they might have a reach large enough to make them an attractive outlet for politicians. The log is again calculated as (log of TV stations +1).

69Indeed, according to the National Association of Broadcasters, the total audience of low power stations including Class A is 800,000 nationwide http://www.nab.org.

7.2 Data and estimation approach 112

on how to distribute grants across his state. Thus, when focusing on the governor’s actions, grant spending in, say county A and B of the same state cannot be the outcome of strategic interaction. It will rather be decided upon by a single person in order to maximize the number of votes, taking into account voters’ reactions to a variation in the allocation of grants. This implies that the Governor’s decisions are quite likely to be driven to a large extent by measurable population characteristics rather than policy interdependence.⁷⁰ However, there may be other political agents such as the aforementioned congressmen or elected county officials that make strategic interaction seem rather conceivable. In addition, such strategic interaction can of course occur in counties bordering another state.

We already try to account for this fact by including the variable DMA home share in our estimations. This variable measures the percentage of the population in a county’s DMA living in the same state the county under consideration belongs to. The higher the share living outside the home state, the more information about what is going on in the other state we expect TV stations to convey, thus creating yardstick competition among Governors. Even though we include this control and we do not feel the spatial dependence in our setting to be an exclusively strategic one, in order to account for the above mentioned effects, we estimate a spatial lag regression model which can be displayed in matrix form as follows:

g = ρ W g+Xβ+ε, (45)

where ε is a vector of i.i.d. error terms, g is a vector representing grant spending, W is a spatial weight matrix,βis a vector of coefficients to be estimated andW ggives the measure of grant spending in neighboring counties. The interaction between own and neighbors’

spending is captured in the coefficient to be estimated, ρ, which we would then expect to have a positive sign. Another reason for the choice of the spatial lag model could be spillovers which we may not be able to capture in the baseline specification. In this case, the spatial correlation, as expressed in ρ may point in either direction.

70If people are envious of the amount of grants their neighboring counties receive, a sort of interdependence would be introduced in that the Governor cannot distribute his funds unequally but must rather follow up on a grant award to county A with an award to county B, thus creating positive spatial autocorrelation.

7.2 Data and estimation approach 113

Another rationale for spatial correlation in our context would be locally correlated shocks or the existence of spatially correlated omitted variables which drive the governor’s choice of local spending. In both cases the spatial interdependence is relegated to the error term, yielding the following spatial error model appropriate:

g = Xβ+ε (46)

ε = λ W ε+u, (47)

where the notation differs from above in thatεis a vector of spatially autocorrelated error terms, u is a vector of i.i.d. error terms and λ is the parameter measuring the extent of spatial autocorrelation. We also estimate a specification that allows for the simultaneous presence of spatial lag and error. Essentially this means estimating equation (45), where the error term is as in equation (47), via a three step procedure that takes into account the endogeneity of the spatially lagged variable.⁷¹

It must be pointed out that these models will be estimated as a robustness check rather than as a means of determining what mechanism is responsible for possible spatial dependencies.

Our interest is mainly in determining whether the main media related variablesdistance to media city and log number of TV stations pick up some of the spatial effects and whether standard errors may be biased downwards in the OLS specification due to the neglect of spatial effects.

7.2.4 Data Sources

The data mentioned above is gathered from a variety of sources. While the dependent variablefederal grants per capitais taken from theConsolidated Federal Funds Report 2000

71Lag and error specification are estimated using maximum likelihood (ML), the combined spatial lag and error model via the GS2SLS estimator proposed by Kelejian and Prucha (1998). The weighting matrix W is row standardized based on rook contiguity, i.e. counties sharing a common border are treated as neighbors.

7.3 Empirical Results 114

(CFFR), many sociodemographic controls stem from theCounty Databook 2000, published by the US Census Bureau. More sociodemographic controls are taken from the database County Profiles published by the US Department of Agriculture.⁷² The county distances to the nearest media city are calculated based on the county population centroids provided by the Census Bureau and the geographic location of media cities obtained by using geocoding software. The names of DMAs and the media cities are those defined by Nielsen Media Research for the year 2002. Counties are assigned to DMAs based on the Nielsen definitions of the same year. The number of fully-licensed as well as low-power TV stations by county is calculated using the Federal Bureau of Communications’ Wireless Telecommunications Bureau Database as of July 2006. Unfortunately, we were unable to obtain data for the actual time period under consideration, yet we believe that given the little variation in the data over time mentioned above, this does not hurt our results too much. The number of votes cast for the Republican and Democratic parties in the presidential elections from 1980 to 1996 is taken from the USA Counties 1998 CD published by the US Census Bureau.

This data was combined with the intercensal population estimates (provided by the same source) in order to calculate vote shares of the Republican and Democrat parties in the presidential elections as well as voter turnout and voter mobility (density). Finally, the number of television sets by county in 1960 is taken from the ICPSR County and City Data Book Consolidated File: County Data 1947-1977. Micropolitan areas are as of 2003, because this classification did not yet exist in 2000. Summary statistics are displayed in table 7–1

7.3 Empirical Results